Open to internships and roles
ResumeSystems write-up with source and docs. Not a visual product case.
A multi-agent system for finding opportunities someone is qualified for, preparing application materials from verified evidence, pausing before sensitive actions, and recording the final result.
AI
Backend
Python
GitHub
Claude
Codex
Cursor
CopilotThe system starts with a person's evidence, such as a resume, transcript, project notes, or other proof. It uses that information to build a profile, search for opportunities that match the evidence, prepare materials, and save the result of each step.
The key piece is ARMCL, a three-level memory loop. It keeps live step data, current task history, and long-term facts separate so the system can recover IDs, approvals, past failures, and hard requirements without asking the same questions again.
Applicant evidence, opportunity requirements, and institutional rules stay separate so a user document cannot quietly change the decision rules.
The system can recommend and prepare materials, but external submission waits for approval unless the user already gave clear permission.
When a plan is rejected or unsafe, the system records that result so a later run does not try the same path again.
A collaborative intake partner screens uploaded evidence, searches only inside the correct tenant and application, identifies missing proof, and prepares a handoff. A separate opportunity partner turns that evidence into a grounded profile, finds matches, and prepares materials without owning the submission tool.
The autonomous fleet then runs through scout, analyst, approval, executor, critic, and explainer roles. Each role has a tighter job than a general assistant: scout finds the case, analyst checks evidence and policy, approval pauses risky actions, executor records the action, critic verifies the artifact or receipt, and explainer reconstructs the decision afterward.
This separation is the point of the project. It shows how I think about agent systems in Python when the model is allowed to help with real decisions: narrow responsibilities, explicit memory permissions, deterministic checks, and a visible audit trail.
I designed the ARMCL loop around scratchpad, episodic, and semantic memory so the system can recover IDs, constraints, approvals, and previous failures without asking the user again.
The fleet separates intake, search, action, verification, and explanation into different workers with declared capabilities instead of giving one model every tool.
The project includes approval gates, idempotency keys, circuit-breaker behavior, redaction boundaries, and refusal records that survive later runs.
The local run path prints node transitions and memory state, so someone reviewing the repo can see the workflow without deploying it.
Fluffy Memory is not just "AI for job applications." The stronger story is that I designed a workflow where evidence cannot silently become policy, model output cannot submit without authorization, and the final outcome can be inspected later.
The point is product value and backend risk at the same time. The user wants opportunities, but the system still has to protect truthfulness, consent, privacy, and repeatability.
The repo includes the system write-up, agent responsibilities, local run instructions, and offline deterministic tests.